WE-DE-207B-03: Influence of Local Anatomical Variations On Detection of Multifocal and Multicentric Breast Cancer.

PURPOSE: We employed a novel model observer to assess the impact of local anatomic variations on the detection of multiple breast tumors depicted on mammograms. We expect the study to be valuable for future task-based assessments and optimizations of x-ray based breast imaging techniques (e.g., digital breast tomosynthesis for diagnosis of multifocal multicentric breast cancer (MMBC)).

METHODS: Regions of interest (ROIs) from four different sets of clustered lumpy background simulations were extracted as the image background. A random number of circular Gaussians were inserted to simulate cases with multicentric lesions. The task of the model observer was to perform a multiple-lesion detection task, making both image-level and location-specific detection decisions based on the overall information from individual ROIs and their interactions. The detectability at each of the possible signal location was measured using channelized Hotelling observers with a novel implementation of partial least squares channels. The power law exponent β of local anatomical noise power spectrum, and spatial covariances K within the ROIs were computed to represent local background variations. The relationships between β, K and the detectability were evaluated.

RESULTS: β and/or K were different across the ROIs. Location-specific signal-to-noise ratio (SNR) results showed statistically significant correlations between β, K and the detectability. Specifically, a ROI with a higher value of β, and/or K with larger and more variable off-diagonal elements was associated with a lower SNR at that ROI. These results also demonstrate the potential of the model observer to adapt to the variations in anatomical backgrounds.

CONCLUSION: We conducted a model observer study to explore the influence of local anatomical backgrounds on detecting MMBC. The study suggests that the optimal imaging settings may need to be adjusted when the clinical task of interest changes (e.g., find only the largest tumor versus find every tumor). This work is supported by National Science Foundation (NSF) Award CBET-1445713.